Llama 3.1 Converse calls not properly escaping quotes in tool calls

0

It looks like tool calls with Llama 3.1 models are not being escaped properly. When the model uses quotes, the tool call params are truncated. This doesn't seem to happen with another model, like Mistral Large.

$ aws bedrock-runtime converse \
  --model-id 'meta.llama3-1-405b-instruct-v1:0' \
  --messages='[
    {
      "role": "user",
      "content": [
        {
          "text": "run this query: select * from users where status=\"active\""
        }
      ]
    }
  ]' \
  --tool-config='{
    "tools": [
      {
        "toolSpec": {
          "name": "run_query",
          "description": "run a SQL query against a DB",
          "inputSchema": {
            "json": {
              "type": "object",
              "properties": {
                "query": {
                  "type": "string",
                  "description": "sql query"
                }
              }
            }
          }
        }}]}'

Llama 3.1 output:

{
  "output": {
      "message": {
          "role": "assistant",
          "content": [
              {
                  "toolUse": {
                      "toolUseId": "tooluse_S4ox6qc9QcCS-iK6AKW3Cw",
                      "name": "run_query",
                      "input": {
                          "query": "select * from users where status="
                      }
                  }
              }
          ]
      }
  },
  "stopReason": "tool_use",
  "usage": {
      "inputTokens": 70,
      "outputTokens": 25,
      "totalTokens": 95
  },
  "metrics": {
      "latencyMs": 1194
  }
}

Notice how the query is truncated. If I change the prompt to

run this query: select * from users where status=\"active\". Replace quotes with \\u0027

I sometimes correctly get

"query": "select * from users where status='active'"

But switching to a Mistral model mistral.mistral-large-2407-v1:0 gives the correct query without prompting to escape.

aws bedrock-runtime converse \
  --model-id 'mistral.mistral-large-2407-v1:0' \
  --messages='[
    {
      "role": "user",
      "content": [
        {
          "text": "run this query: select * from users where status=\"active\""
        }
      ]
    }
  ]' \
  --tool-config='{
    "tools": [
      {
        "toolSpec": {
          "name": "run_query",
          "description": "run a SQL query against a DB",
          "inputSchema": {
            "json": {
              "type": "object",
              "properties": {
                "query": {
                  "type": "string",
                  "description": "sql query"
                }
              }
            }
          }
        }}]}'

Mistral output:

{
  "output": {
      "message": {
          "role": "assistant",
          "content": [
              {
                  "toolUse": {
                      "toolUseId": "tooluse_WUcMjafeRRGtAfGUmPNWRA",
                      "name": "run_query",
                      "input": {
                          "query": "select * from users where status=\"active\""
                      }
                  }
              }
          ]
      }
  },
  "stopReason": "tool_use",
  "usage": {
      "inputTokens": 79,
      "outputTokens": 30,
      "totalTokens": 109
  },
  "metrics": {
      "latencyMs": 1924
  }
}
asked 3 months ago297 views
1 Answer
0

Hi,

What you describe really looks like a bug. Service teams do not monitor re:Post for bug reports.

So, the best is to open a support ticket with your precise description above via the console of your AWS account.

Best,

Didier

profile pictureAWS
EXPERT
answered 3 months ago

You are not logged in. Log in to post an answer.

A good answer clearly answers the question and provides constructive feedback and encourages professional growth in the question asker.

Guidelines for Answering Questions